P. Marjai, A. Kiss: A Survey on the Usage of Centrality Measures in Error Prediction. In Intelligent Computing (Lecture Notes in Networks and Systems 1019). Cham, CH : Springer, pp. 157–167, 2024. ISSN 2367-3370, ISBN 978-3-031-62273-1 link

Abstract: The advent of extensive systems has led to a swift escalation in the volume of generated logs. Virtually every software application generates these files, which encompass runtime data of the software, like significant occurrences or potentially concerning behaviors such as errors. Log files serve as a valuable reservoir of information for comprehending and overseeing system operations, enabling the anticipation of forthcoming irregularities. In recent times, several approaches have emerged to address this objective. These include both supervised and unsupervised models, alongside deep learning techniques. Since error prediction is critical in fields like reducing downtime and performance optimization, root cause analysis and security monitoring, our goal is to examine new correlations between the lines of the log file that can help predict upcoming errors and aid the software operators in the above-mentioned tasks. Because of this, in this paper, we investigate the prediction capacity of a new heuristic method that builds a precedence graph based on the log templates that follow each other. The templates also called events are acquired with the use of a template miner. These events are then used to create the precedence graph. After the creation of the graph, different centrality measures are employed to identify the errors. We conclude that there is a connection between a node’s anomaly detection ability and it’s well-connectedness, while there is no connection between the node’s position on shortest paths and it’s ability to predict anomalous behaviors.